基于图强化学习的含光储配电网无功优化方法  

Reactive Power Optimization Method for Distribution Networks with Photovoltaic and Energy Storage Based on Multi-agent Graph Reinforcement Learning

作　　者：全欢 胡建军 刘肇熙 谭科 祝敬华 陈逸凡 肖嘉睿 QUAN Huan;HU Jianjun;LIU Zhaoxi;TANG Ke;ZHU Jinghua;CHEN Yifan;XIAO Jiarui(School of Electric Power Engineering,South China University of Technology,Guangzhou,Guangdong 510641,China;Management Science Research Institute,Guangdong Power Grid Co.,Ltd.,Guangzhou,Guangdong 510699,China)

机构地区：[1]华南理工大学电力学院,广东广州510641 [2]广东电网有限责任公司管理科学研究院,广东广州510699

出　　处：《广东电力》2024年第12期79-86,共8页Guangdong Electric Power

基　　金：广东省基础与应用基础研究基金项目(2023A1515011171);广东省新型电力系统技术创新研究项目(1690186045357)。

摘　　要：传统的无功优化方法已无法适应大规模分布式能源接入下的新型配电网,为了更好的应对含高比例新能源配电网的无功实时调度,提出一种基于图强化学习的配电网无功优化方法。首先,在分区分散式控制框架下根据配电网无功优化控制模型构造部分可观测马尔可夫决策过程。然后,提出基于物理信息强化学习范式的多智能体深度确定性策略梯度算法,通过图卷积神经网络挖掘拓扑结构的物理信息。随后,基于集中式训练-分散式执行的框架对马尔可夫决策过程进行求解,得到离线训练阶段的最优策略。最后,将训练好的模型在线应用,仿真结果表明所提方法能够有效应对电压波动和源荷不确定性,具有较好的实时性、泛化性和无功优化效果。Traditional reactive power optimization methods have been unable to adapt to the new distribution network under the access of large-scale distributed generation.In order to better cope with the real-time reactive power scheduling of the distribution network with a high proportion of new energy,this paper proposes an graph-mechanism based reactive power optimization reinforcement learning method for the distribution network.Firstly,a partially observable Markov decision process is constructed according to the reactive power optimal control model of distribution network under the framework of zonal decentralized control.Then,a multi-agent deep deterministic policy gradient algorithm based on graph convolutional network is proposed to mine the physical information of topology structure through graph convolutional neural network.Afterwards,based on the framework of centralized training and distributed execution,the Markov decision process is solved,and the optimal strategy of off-line training is obtained.Finally,the trained model is applied online.The simulation results show that the proposed method can effectively deal with voltage fluctuation and source load uncertainty,and has good real-time and reactive power optimization effect when applied to large-scale systems.

关 键 词：无功优化 多智能体图强化学习 马尔可夫决策过程 图卷积网络 配电网 

分 类 号：TM711[电气工程—电力系统及自动化]